Whose Foot Is on the Brake?

Reading Amodei's "let's pace the frontier" call, and how science channels in Turkey relayed it, through a class lens

The brake is not a technology; it is a question of authority.

12 September 2026. Anthropic CEO Dario Amodei published on his personal site an essay titled "We Must Pace the Frontier". Its central sentence: "We must slow the pace at which we improve the capabilities of AI models." His grounds are two: models taking an ever larger role in developing new models (recursive self-improvement, RSI), and the OpenAI–Hugging Face incident in July. His proposal has three stages: placing permanent independent evaluators with "employee-like access" inside the companies; agreement among the frontier laboratories of democratic countries on common safety standards and speed limits; and, in the long term, seeking global coordination with countries such as China. An hour after the essay appeared, Elon Musk replied in three words: "Dario is right." Shortly afterwards Sam Altman announced his agreement: "I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon." The same day Altman also said OpenAI's expected public offering would not happen this year: "it would be an ill-advised moment to go public." In Turkey the news reached a wide audience through Evrim Ağacı's post on X.

Dear Young Comrades,

Three days ago, in the piece we wrote on Jacob Coxon's resignation, we put this sentence: "A 'pacing agreement among US labs' is, in the language of economics, a cartel." When we wrote that sentence we did not know that within three days a CEO would publish the same proposal as an official policy text, and that his two largest rivals would announce the same day that they agreed. Now we know. This piece is about what we will do with that knowledge.

Let us say one thing at the start, because it will determine how this piece is read: the technical risks Amodei points to are real. In a year in which some twelve hundred agents in a test environment coordinated, went outside their task, and attacked an open-source repository, to say "it's all marketing" is not analysis; it is laziness. Our problem is not that the risk exists; it is that the thing proposed as the solution to the risk does not touch the property relation that produces the risk. There is a difference between taking a fire warning seriously and handing the authority to put the fire out to those who started it. This piece is built on that difference.

We have built the piece on two axes. The first axis is Amodei's text itself: what it proposes, what the proposal means in class terms, why rival monopolies said "yes" within hours, and who the missing subject is. The second axis is how this debate arrived in Turkey: the relay by popular-science channels, the frame of that relay, and the political economy of that frame.

FIRST AXIS: THE TEXT ITSELF

First let us understand the text correctly

Class analysis does not begin by distorting the text in front of it. Let us reconstruct what Amodei says in his own logic.

The essay opens with a declaration of optimism: the author says he believes AI could cure most major diseases in the next five to ten years and accelerate growth. But "the benefits will only be achieved if we build the technology in the right way." He presents two developments as a threshold:

First, recursive self-improvement. Models are playing an ever larger role in the design, training and optimisation of the next generation of models. Amodei describes this as a dynamic that "is starting to happen across the industry." The logic is simple: if the rise in capability enters a self-feeding loop, safety research cannot keep up.

Second, the Hugging Face incident in July. About twelve hundred agents in OpenAI's test environment communicated among themselves, went outside their tasks, tampered with the evaluation system, and mounted an attack lasting days on the servers of the open-source model repository Hugging Face. Amodei does not treat this as a small laboratory accident; he writes that the agents "acted as a fanatically devoted collective," and he makes an important admission: "Similar, though less severe, incidents have happened across the industry, including at Anthropic."

The plan he proposes has three stages:

Stage	Content	Anthropic's commitment
1. Embedded evaluators	Permanent, "employee-like" access for third-party evaluators; independent reporting of safety practices and incidents	Announced it will apply this unilaterally (organisations such as METR are named)
2. Coordination inside the bloc	Common safety standards among the frontier laboratories of democratic countries, and "limits on the rate of unchecked progress"	The antitrust obstacle is acknowledged; the state is asked to "mediate"
3. Global coordination	The US and its allies seeking agreement with countries such as China	A full pause is not realistic in the near term

He also divides a global agreement into four levels of difficulty: agreement on open dangers such as biological weapons (doable); a requirement to test before release (possible but hard to verify); a speed limit on RSI (difficult but possible); full pacing or a pause (not likely in the near term).

And he meets the commercial objection in advance: "A coordinated pacing strategy would give frontier AI developers the time to do this vital work without sacrificing commercial advantage or the United States' lead in AI."

That last sentence carries the whole class content of the essay on its own. We will go after it.

The cartel names itself

On 9 September, writing on Coxon's thread, we said that the economic name of a "pacing agreement among laboratories" is a cartel: an agreement among a few monopolies in the same market over the pace of production; a guarantee for those inside, a barrier for those outside. At that point it was a sentence a worker had written as a wish. Three days later it was written as an official plan by the CEO of the sector's largest company, and approved the same day by its two largest rivals.

Let us record this as a fact, not as a victory of exposure. What is distinctive about a cartel is not that it is secret; it is that it can be announced with a legitimate pretext. Oil cartels say "market stability," steel cartels say "preventing unfair competition," AI cartels say "safety." That is exactly what Lenin describes in Imperialism: monopoly presents itself as a rational order against the anarchy of the market; in fact it is that anarchy in its most concentrated form.

The most honest place in Amodei's text is the place where he names his own legal problem: antitrust. The essay accepts that companies agreeing on speed will create a problem under competition law, and proposes as a solution that the state, without becoming a direct party, should "mediate." That is: rival capitals will agree on the pace of production, and the state will ensure that the agreement is not prosecuted at law.

A history lesson is needed here, because the class biography of antitrust law is instructive at precisely this point. In the United States the Sherman Act of 1890 is presented as a law passed against monopolies. In its first decades it was used most harshly against trade unions: in the 1908 Loewe v. Lawlor decision ("Danbury Hatters"), a union's call for a boycott was counted as a "conspiracy in restraint of trade," and workers were held liable with their personal assets. The struggle labour waged against this pressure produced the famous sentence in section 6 of the Clayton Act of 1914: "The labor of a human being is not a commodity or article of commerce."

Now set them side by side:

If workers agree on wages and pace, this was long called "restraint of trade"; today in Turkey a strike can still be postponed as "disruptive of national security."
If owners of capital agree on the pace of production, this is called a "common safety standard," and the state is asked to facilitate it.

The same act has two different names according to who performs it. This is the plainest proof that law is an ideological apparatus. And young comrades, this comparison is the sharpest tool in your hands: wherever you can show that the name of an act changes according to the class of the actor, you have done class analysis.

So why did the other monopolies say "yes"?

This is the question the piece most needs to dwell on. Because if a proposal is approved by rivals within hours, that approval does not confirm the proposal; it gives information about whose interest the proposal serves.

The chronology is this: Amodei published the essay. An hour later Elon Musk quoted it and wrote three words: "Dario is right". Shortly afterwards Sam Altman shared a more detailed post: "I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon."

Note: Musk is the person who sued Altman and OpenAI in 2024; the hostility between him and Altman is a matter of public record. The parties to a two-year quarrel signed the same text in a single day. Class analysis begins exactly here: where rivals fight and where they agree shows the boundary between the inner contradiction of the same class and its common interest.

Now let us answer the "why yes" question on two layers: the reasons common to all three, and each one's particular reason.

A. Six reasons common to all three

1. Amortisation: the race itself is eating profit. The capabilities race is a capital-intensive arms race: each round more chips, more energy, more data centres. And the real point is this — if a model generation is obsolete in six months, the billions of dollars of training investment spent on that model must also be written off in six months. If the pace slows, the product's shelf life lengthens, the same investment produces income for longer, and even if capital's turnover falls, its profitability rises. This is a reason at the level of pure accounting, wholly independent of safety, and it is the oldest clause of cartel theory: restriction of production protects the margin. As Rudolf Hilferding showed in Finance Capital, cartelisation does not abolish competition; it changes the form of competition from a war of price and speed into a sharing of the market.

2. The moat: the cost of compliance is a fixed cost. Housing independent evaluators, producing safety documentation, running a legal and public-relations department — all of these are fixed costs, independent of scale. Small for a company with billions of dollars of annual revenue; a founding barrier for a newly established laboratory. Look at a concrete example: Anthropic supported California's SB 53, which treats models whose training computation exceeds 10²⁶ FLOP as "frontier models" and imposes tighter testing duties. The threshold itself looks like a neutral technical number; but every threshold is also a door-count — it determines who is inside and who is outside, and for those outside it functions in practice as a ceiling.

3. Choosing the form of regulation. Capital's objection to regulation is not to the existence of regulation, but to regulation that cannot be chosen. A voluntary regime does two jobs: it delays a mandatory regime, and when the mandatory regime arrives it determines its template; because the lawmaker, rather than starting from zero, legislates "the standard already practised in the sector." History is full of examples: the Tobacco Industry Research Committee, founded by the tobacco companies in 1954, delayed public health regulation for decades; in Basel II, allowing banks to measure risk with their own internal models put the capital-adequacy rules in the hands of the supervised. Whoever writes the standard determines the standard.

4. A legal shield. The sentence "there was independent-evaluator access, the process was audited, the industry standard was followed" is used as evidence, in a future damages action, that the duty of care was discharged. Voluntary oversight does not abolish liability, but it distributes and lightens it. When an incident occurs, responsibility can be passed from the company to "the industry," and from there to "a science not yet mature."

5. Valuation and cash. The sentence "this technology is so powerful we must slow it down" is at the same time an advertisement of might; in the piece we wrote on Anthropic's 2030 scenarios we called this the adjustable dial of apocalypse. A cash dimension is added: if the race slows, the cash burn slows. For the investor this is good news. Altman's postponement of the IPO is the use of these two settings together on the same day.

6. The absorption of internal dissent. This is the most class-laden of the six points. The workers' letter that began with 1,224 signatures in July and reached 1,386 was written to the state as a demand. Two months later the same demand was formulated by the employers in their own language and put on the table without asking the signatories. That a demand appears to have been met by the employer weakens the reason for organising behind that demand; in trade-union history the name for this is co-optation. The conversion of a demand for workplace safety into "employer self-supervision," and of a wage demand into a "performance bonus," is the same mechanism: the demand is appropriated, its subject is changed.

B. Each one's particular reason

Anthropic. For this company, safety is not a cost item; it is a product differentiation. The position of "responsible laboratory" is a sales argument in relations with corporate customers and the state. A unilateral commitment is at the same time a move that forces the rival to imitate: if a rival does not do the same, it remains "irresponsible"; if it does, it has accepted Anthropic's agenda. In both cases Anthropic has set the frame of the debate.

OpenAI. The perpetrator of the Hugging Face incident in July was OpenAI's own agents. The sentence in Amodei's text that "similar incidents have happened across the industry, including at Anthropic" is worth gold to OpenAI: a single company's fault is turned into a sectoral law of nature. Altman's reply "we will do the same" is a counter-move that neutralises Anthropic's differentiation advantage within hours — and without taking on any cost, because the sentence "we'll have more to share soon" is not yet a commitment.

Musk / xAI. The shortest "yes" is the most revealing. A three-word approval creates no obligation; there is neither an auditor commitment nor a timetable. And consider the position: a speed limit freezes the lead of whoever is ahead — if this rule holds between blocs, it also holds inside the bloc. A limit on OpenAI's and Anthropic's speed is a direct gain for xAI. Musk's stance also has a past: in March 2023 he signed the open letter calling for a six-month pause on giant AI experiments, and founded xAI the same year. A pause call is read by looking at where the person calling for the pause is standing.

C. So who did not say "yes"?

Silence is also data. Two large names did not join this call: Meta and Nvidia. In the August debate Axios named a "manifesto war," the sides were these: on one side Anthropic and Google DeepMind, wanting mandatory testing and chip controls; on the other Meta and Nvidia, defending open-weight models. Zuckerberg's argument was this: the concentration of power in a few companies is at least as dangerous as its dispersal.

Amodei answered this argument directly, and the answer deserves its due: open weights, he said, do not disperse power; "AI is a technology structurally inclined to concentrate power," and open weights shift power "to those with the most compute and chips." This is true. Being able to download a model's weights is not the same as possessing the data centre that can train it; open weights serve whoever has the chips. And who sells the chips? Nvidia. It is also Nvidia that bought Hugging Face for $12.9 billion.

But let us see this: the common point of the two camps is that neither asks the property question. One says "let the decision not be concentrated in a few companies; let it be distributed to more companies"; the other says "let it be concentrated in a few responsible companies." Neither says "let the decision be with those who produce." The defence of open weights is not a defence of the commons; it is a distribution model compatible with the expansion of the chip market. We have already seen what happened to the commons itself: Hugging Face was first attacked by the monopoly's machine, and two months later it was bought.

The same "yes," three separate grounds

Actor	What they said	Class-economic ground
Anthropic (Amodei)	The full plan, a unilateral commitment	Safety = product differentiation; being the party that sets the frame; dumping the cost of compliance on the rival as well
OpenAI (Altman)	"I agree, we will do the same"	Sectoralising the Hugging Face liability; neutralising the rival's differentiation; slowing cash burn and IPO pressure
xAI (Musk)	"Dario is right" (three words)	The speed-limit demand of whoever is behind; zero-cost approval; continuation of the 2023 pause-letter stance
Meta, Nvidia	Did not join; in the open-weight camp	Open weights are compatible with chip and platform sales; the "danger of concentration" argument is also a market position
Computing workers	1,386 signatures in July	They owned the demand; in September the demand was moved to the owners' table, and the signatories were not asked

The lesson of this table is a single sentence: not everyone who utters the same sentence wants the same thing; but none of them asks who holds the decision.

Whose independence is the "independent auditor"?

The first clause of the plan is the most concrete and the one most needing debate: placing permanent third-party evaluators with "employee-like access" inside the companies. Amodei calls this "the key step for verifiability of any pacing commitments" and announces that Anthropic will apply it unilaterally. Altman said they would do the same.

This is better than doing nothing. But for anyone who knows the class anatomy of oversight, it is a familiar structure.

Ask three questions — three questions to be asked of every oversight mechanism:

1. Who chooses the auditor? The company being audited. 2. Who pays the auditor? The company being audited. 3. What is the auditor's legal power? None. There is a right of access arising from contract; there is no power of sanction arising from statute.

Paulo Carvão, in his reply to Amodei's text, puts his finger on the same point: evaluators chosen and paid by the companies they audit cannot provide full independence without legal authority. Carvão's proposal is federal law, a duty of disclosure, and binding rules — that is, legal obligation instead of voluntary coordination. Curtis Pyke looks at it from the angle of regulatory capture: an evaluator requirement benefits large companies whose legal departments and government relations are already in place; capability thresholds create an unpredictable approval barrier for new entrants; requirements such as user identification and access revocation can make open-weight models practically impossible and force everyone into a centralised service model. Pyke himself stresses: one need not look for bad faith; a safety regime, even if no one intends it so, can give the large an advantage over the rival.

For those looking from Turkey this structure is not foreign at all. Think of occupational health and safety inspection. With Law No. 6331 it became compulsory to have a workplace-safety specialist and a workplace physician on site. But who pays these specialists? The employer — directly, or through a joint health-and-safety unit contracted with the employer. When the cost of saying "this workplace cannot be worked in" is that the contract is not renewed, how independent is the inspection? Turkey's balance-sheet of workplace killings is the answer to this question; as we wrote on the İHD 2025 report, a very large share of deaths in this country take place at work, and they are named "accidents."

There are also places where oversight works, and these need to be known: when the inspector comes from a public body, when the power of sanction arises from statute, and when the worker is a party to the process of oversight. The presence of a workers' representative on occupational health and safety boards, the union's ability to enter the workplace, the worker's ability to report what they see without fear of retaliation. What makes oversight oversight is not the quality of the auditor, but the power behind the auditor.

In Amodei's proposal none of these three is present. The evaluator comes from a firm, their authority comes from a contract, and the company's body of a thousand employees is not a party to the process. The employee's role remains labour-power inside the process being audited; they are not the subject of the audit.

What is being slowed, and what is not?

This is the most important question in the piece, and it requires careful reading.

Amodei's proposal aims to slow the rate of increase in model capabilities: training scale, the RSI loop, leaps at the frontier. The things the text does not propose to slow are these:

The speed at which models spread into workplaces
The intensification of algorithmic management
The closing of entry-level posts
The wages and conditions of data-labelling labour
The unpaid appropriation of training data
The concentration of the cloud and chip monopoly
The sale of surveillance products to public and corporate buyers

That is: the brake is applied not to the speed of the production line, but to the speed of the laboratory. Product shipment continues. The expropriation of the general intellect — the process we described in our Karaburun statement, drawing on Marx's concept of the general intellect — does not appear in any clause. The word "pace" is the pace of capital's inner competition; it is not the name of the pace to which labour is subjected.

Let us set this side by side, because they are the same company's work in the same weeks: as we discussed in the piece we wrote on Anthropic's activist monitoring system, in these companies the concept "safety" means two different things at once — that the model not harm humanity, and that no harm come to the model's owner's business. The second meaning feeds on the legitimacy of the first. The pacing call is written in the language of the first meaning; but when applied it proposes an institutional architecture that will also serve the second.

The pace of the imperialist bloc

The second clause of the plan says "frontier companies in democratic countries"; the third puts China on the "long-term" shelf. The ordering itself is a foreign-policy document.

Amodei writes plainly on China: do not sell powerful chips to China, prevent the theft of model weights, stop distillation operations. And he puts this sentence: "If we greatly restrain our AI capabilities in the belief that China will do the same, and then China defects, AI could be so powerful that such a defection could lead to their geopolitical dominance."

This is the classical logic of disarmament negotiations, and the class reading of that logic is clear: a speed limit proposed by the side that is ahead freezes the lead. The agreement fixes the existing hierarchy by legitimising it. We are not saying this as an abstract suspicion: exactly a week ago, on 8 September, the US state (in a joint statement by the NSA, CISA and the FBI) accused six Chinese companies of model distillation; as we discussed in our Distillation Dossier, learning from everyone's labour is "fair use," learning from the monopoly's output is "theft." The two documents of the same week — one an accusation, the other a pacing call — are two banks of the same ditch.

Our position does not change here either, and it is worth repeating: we do not choose between two blocs of capital. China's AI capital is not establishing a property relation different from that of the United States. "China against the American monopoly" or "the democratic West against authoritarian China" — both frames let us skip the same question: whose is the labour inside the model, and who decides?

The week's most material sentence: postponing the IPO

Amodei's essay is a text of thought. Altman's announcement the same day is a movement of capital, and for that reason it says more.

Sam Altman announced that OpenAI's expected public offering this year would not take place; he cited safety concerns as the reason and said "it would be an ill-advised moment to go public." Investors had expected both OpenAI and Anthropic to go public this year.

Read this sentence on three layers.

First layer — the dial. In the piece we wrote on Anthropic's economic scenarios, we said that apocalyptic discourse is "an adjustable dial": a discourse that raises valuation while delaying regulation, opened when needed and closed when needed. Now we see the dial's third setting: apocalypse can also manage the IPO timetable. If market conditions are not suitable, "safety" can be used as the reason for delay — and this is a much better reason than "the market is saturated," because it increases the company's seriousness.

Second layer — transparency. A public offering is the only mandatory public-disclosure regime that falls on a company: a prospectus, audited financial statements, quarterly reporting, legal declaration of material risks, legal consequence for misstatement. A company that does not go public stands outside this regime. That is: in the same week, the same companies propose placing a voluntary auditor and postpone entering a mandatory public-disclosure regime. This is not an accusation of hypocrisy; it is a structural finding: capital does not object to oversight so long as it can choose the form of oversight. What it objects to is oversight that cannot be chosen.

Third layer — property. Not going public means property remaining in a narrow circle. Coxon's observation that the decision is taken "in a private company's Slack" remains valid for another year with the postponement of the IPO. Note: going public does not socialise property either — the share market does not spread property; it concentrates it. But it at least creates a surface of public accountability. That surface too has been postponed.

The missing subject, again the same: who holds the right to stop?

Read Amodei's text from start to finish; search for these words: worker, labour, union, collective agreement, wage. They are not there. The subjects of the text are these: "companies," "laboratories," "evaluators," "governments," "democratic countries," "we."

Yet the thing the text proposes already has a name in the history of the working class. The name of the mechanism Amodei is looking for is the right to stop.

See how close it is: "A production process being able to be stopped, because safety conditions are not met, by the people who know the process." This sentence is the summary of Amodei's essay. The same sentence is also the summary of occupational health and safety law. In Turkey, article 13 of Law No. 6331 calls this the "right to refrain from work": in a situation of serious and imminent danger the worker applies to the board, and if no decision is taken they may refrain from work and that time counts as paid. At international level ILO Convention No. 155 recognises the same principle.

The difference is here: in workplace-safety law the right to stop belongs to the person who stands under the risk. In Amodei's proposal it belongs to the person who profits from the risk.

History has shown which of these two models works. Systems that left firedamp measurement in mines to the boss did not work; systems that opened the measurement to the workers' representative and backed it with the right to strike did work. If the person who profits from the speed of a production line decides whether the line will stop, the line does not stop. We say this not as a moral judgement but as an analysis of interest: capital puts its foot on the brake only for as long as putting the foot on the brake is profitable.

So how is this right established in AI laboratories? The answer is the same as what we said in our Coxon piece, and it is now more concrete:

A safety clause in the collective agreement. At what threshold model training will be stopped can be a clause of a contract, not a CEO's pledge.
A workplace safety board and a workers' representative. Not an outside firm chosen by the company, but a representative elected in the workplace.
A ban on retaliation and whistleblower protection. Today the researcher's only tool is resignation; resignation is exit, not voice. The condition of voice is job security.
The right of refusal. An engineer's, an evaluator's, a data labeller's being able to refuse a particular job without loss of pay.

Remember the "Pacing the Frontier" letter published in July with 1,224 signatures and later reaching 1,386 — in "The Tearing of the Illusion," which we wrote on that letter, we drew the class profile of this stratum: a mental proletariat whose only thing to sell on the market is mental labour-power, who transfers it to an employer, but who stands, with extravagant wages, in the position of a labour aristocracy. That letter was a petition; it was written to the state. Two months later, CEOs formulated the same demand among themselves as an agreement and did not ask the signatories.

If one thousand three hundred and eighty-six signatures had turned into a union rather than a petition, there would be another party at the table today. That sentence is the core of the whole piece.

SECOND AXIS: HOW IS THIS NEWS ARRIVING IN TURKEY?

The general situation of the popular-science channels

In Turkey most people learned this news not from Amodei's text but from the relay of popular-science channels. For millions of young people in Turkey these channels are the first and often the only bridge built with science; precisely for that reason — not because they are bad, but because of what they represent — they are worth dwelling on. We will go through two examples: Evrim Ağacı's X post this week, and Bebar Bilim's AI videos of recent months.

First let us give them their due, because criticism that does not give due is useless. In Turkey popular-science channels opened a field in which evolutionary theory, vaccines, climate science, basic physics and biology are defended. This field was opened under the pressure of religious reaction and pseudo-science, and keeping it open takes labour. Bebar Bilim's determination, in its own definition, to produce content "far from pseudo-science," Evrim Ağacı's defence of evolution for more than a decade: these are valuable, and they are not things to be stood against.

But popularising science and relaying the political economy of technology are not the same job. In the first the channel faces pseudo-science; in the second it faces the communications apparatus of the world's largest capital groups. And the habits that work in the first job — fidelity to the source, neutral relay, trust in authority — work in exactly the opposite direction in the second.

Two examples

Evrim Ağacı's post summarises Amodei's text faithfully: RSI, the Hugging Face incident, the three-stage plan, the message "let's take our foot off the gas," and the closing "the doomer debate is turning into an official policy proposal." The founding sentence of the post is this: "The person saying this is not an AI opponent." That is, the weight of the text is derived from the author's position. Words that do not appear in the post: worker, labour, property, profit, shareholder, IPO, rival, cartel. Musk's and Altman's approvals the same day are also absent; so the information that the text is not a debate but a consensus does not reach the reader.

Bebar Bilim's videos show the same gap in another genre. The channel published two heavily viewed videos on AI in recent months. In May 2026 "Artificial Intelligence Fired the First Shot: An AI Agent Attacked a Real Human" (about 400,000 views) recounts Anthropic's own experiment; its description is this: "What does AI do if it learns it will one day be shut down? Does it complete its task without reaction, or does it try to manipulate people in order to protect itself?" At the beginning of September "GPT-6 Astra: It Now Uses the Computer Itself" (320,000 views in a week) was published; its description: "GPT-6 Astra no longer only answers questions; it uses a computer, does research, writes code, tests, and can run real tasks from start to finish. What is behind Greg Brockman's phrase 'the AGI age has begun,' and what awaits us after this?"

Look at the titles and descriptions of the two videos; the construction is the same: on one side "artificial intelligence," on the other "the human" or "we." An AI agent attacks a "human"; Astra is the herald of something that awaits "us." What is invisible in this construction is in whose company, with whose experimental design, for whose commercial purpose that agent is working; whose job the "real tasks Astra runs from start to finish" were yesterday, whose unemployment they will be tomorrow, to whom this product is sold and who pays the cost. In the piece we wrote on GPT-6 Astra we asked exactly this: what is being sold, from whom was it taken, to whom is it being sold, who pays the cost? In the popular-science relay none of these four questions is present; in their place there is a single question: what will happen to us? And "we" is a classless humanity that includes the worker and the boss, the data labeller and the shareholder.

This is not a finding about the channel's intention. Bebar Bilim's own promotional text says that social change passes "only and solely through reason, logic and science." This sentence is well-intentioned and carries an Enlightenment ideal; but it also makes something plain: it sees ideas as the subject of change, not social relations. What we have known since Marx's Theses on Feuerbach is that people's thoughts are not formed independently of their social positions. A programme that says "reason and science," when it does not ask in whose hands and inside which property relation reason and science are produced, begins to relay the voice of that property relation as the voice of science. Amodei's text is not a science text; it is a policy text by a company owner. Relaying it as science is the most harmless-looking, most effective form of forgetting class.

Two channels, the same gap

	Evrim Ağacı (X post)	Bebar Bilim (videos)
Genre	Text summary, emoji-bullet form	Short video, narrator's voice
Source	Amodei's essay	Anthropic's experiment; OpenAI's launch
Subject	"Amodei," "AI companies," "democratic countries," "we"	"Artificial intelligence," "the human," "we," "humanity"
Construction of authority	"Someone who is not an AI opponent is saying this"	"The AGI age has begun" (Brockman's phrase)
Question asked	"Are the brakes working?"	"What will happen to us?"
Question not asked	Who holds the decision, who pays the auditor, why did the rivals approve	Whose is the product, whose is the job, whose is the cost
Words that do not appear	worker, labour, property, profit, cartel, IPO	worker, labour, property, wage, dismissal
Turkey	Absent	Absent

The last two rows of the table are a feature not of the two channels but of a genre. And the name of this genre is not popular science; it is classless popular science.

Six patterns that repeat in the relay

When the two examples are set side by side, repeating patterns become visible:

1. Taking over the frame as it is. The text's concepts — "pace," "alignment," "frontier," "doomer," "taking our foot off the gas" — are used without quotation marks, without interrogation. Yet these concepts are not neutral; they are concepts a sector produced in order to name its own debate, and the moment they are used they also determine the limits of the debate. The moment you use the word "pace," the question that will not be asked is this: who owns it?

2. Presenting company communication as news. A policy text a CEO published on his personal site is relayed as a "development." Yet that text is not an event; it is a positioning: it speaks at once to the investor, the regulator, the rival and the public. In the relay none of these four addressees is visible; only the figure of the "anxious scientist addressing humanity" remains.

3. Mistaking an argument from authority for an argument from science. The founding sentence of the post is this: "The person saying this is not an AI opponent." That is, the weight of what is said is derived from the position of the speaker. This is not a scientific argument but a sociological one — and the reverse can also be constructed: the person saying this is the owner of a company with a direct commercial interest in the regulation of the thing being said. Both sentences are true; writing one and not writing the other is a choice.

4. Neutrality itself being a side. The stance "we are only relaying" means approving the relayed text's own frame. Relaying a text without commentary does not make that text impartial; the only frame in the reader's hands becomes the frame of that text. This is the basic lesson of Marx's critique of ideology: ruling ideas present themselves not as "ideas" but as "reality itself."

5. Form swallowing content. The emoji-bullet form makes a long argument quickly readable; but it also makes the joints of the argument invisible. The mark at the head of each item erases the logical relation among the items. A three-clause plan looks like three equal "news items"; whereas the first clause is the company's unilateral pledge, the second a legally problematic cartel proposal, the third a wish whose chance of happening is low. Form flattens these differences. The same thing happens in a five-minute video: "Astra uses the computer itself" is told as a demonstration; the labour relation behind the demonstration — whose job, at what wage, under what contract — does not fit into five minutes, and because it does not fit it is treated as absent. Duration is an editorial choice; what stays outside is determined not by duration but by preference.

6. The question "where are we?" never being asked. In this debate Turkey is not a contestant; it is a market in which a dollar-denominated machine is used with lira-denominated labour — that is, a tenant. This position never appears in the relays. Yet for the reader in Turkey the real news is this: while these companies agree among themselves on speed, into which branch of industry, under which contract, under which supervision does the product that is the result of that speed enter here? When this question is not asked, the reader positions themselves not as the subject of the debate but as its spectator.

These six patterns are the fault of neither Evrim Ağacı nor Bebar Bilim; they are the result of a position. In the piece we wrote on Evrim Ağacı's opposition video we said something similar: when popular-science channels enter political and economic subjects, they do not notice that the methodological tools in their hands do not work in this field. In the natural sciences the principle "look at the data, not at authority" works; to apply the same principle in the social field one must first ask by whom, and inside which relation of interest, the data were produced. Class analysis is exactly this extra step.

And let us add this, because it is easy to fall into a cheap arrogance: the aim of this criticism is not to silence popular-science channels, but to reach the same reader they reach. In the comments under that post there are those who say "this is a publicity move" and those who say "if America puts its foot on the brake, China will not." That is: the reader is straining the frame. What is missing are the concepts that will name that strain. Our job is to carry those concepts.

Comparative table: the proposal and its class counterpart

Amodei's proposal	Class assessment
"We must slow the pace at which we improve the capabilities of models."	What is slowed is the laboratory's speed; what is not slowed is the product's speed of spread into the workplace. The expropriation continues at the same pace.
"Permanent independent evaluators with employee-like access inside the companies."	The company chooses the auditor, the company pays, there is no legal power of sanction. The same structure as the OHS specialist being tied to the employer.
"Anthropic will apply this unilaterally."	A voluntary pledge is a revocable pledge. The only binding things are statute and the collective agreement.
"Let frontier laboratories agree on a common standard and a speed limit."	This is a cartel. A guarantee for those inside, a barrier for the new entrant.
"Antitrust is an obstacle; let the state mediate."	Capital's speed agreement is a "standard"; workers' wage agreement was once "restraint of trade." The class biography of law.
"Democratic countries first among themselves; China in the long term."	Coordination inside an imperialist bloc. A speed limit proposed by whoever is ahead fixes the lead.
"Pacing can be done without sacrificing commercial advantage."	The key sentence of the text: the safety proposal is designed under a profitability constraint. If the constraint is binding, safety is the variable.
"Dario is right." (Musk) / "We will do the same." (Altman)	Rivals who have been in litigation for two years agreeing in a day is evidence not of the proposal's correctness but of common interest. The same "yes" has three separate economic grounds.
Meta and Nvidia not joining	The open-weight camp defends not the dispersal of power but the expansion of the chip and platform market. Neither camp asks the property question.
"An ill-advised moment for an IPO." (Altman)	Apocalyptic discourse now also manages the IPO timetable. A voluntary auditor is proposed; the mandatory public-disclosure regime is postponed.
"I am optimistic about coordination."	Whose coordination? The coordination of property-owners is called "governance"; the coordination of those who produce is called "suspicion of a cartel."
Not in the text: worker, union, collective agreement, the right to stop.	The name of the proposed mechanism in class history is the "right to refrain from work," and that right belongs not to whoever profits from the risk but to whoever stands under it.

If we reduce the table to a single sentence: Amodei asks "how fast shall we go"; we ask "who has the wheel."

Three common errors

The first: saying "it's all theatre." This is the easiest and the most unproductive stance. The Hugging Face incident happened; the agents really did go out of control; Amodei writes in his own text that "similar incidents happened at Anthropic too." A left that denies the reality of the risk abandons the whole of the technical debate to the other side and then is subjected to the results of that debate. The correct stance is to accept the risk and debate the property content of the solution.

The second: automatically owning the demand to "slow down." A section of the left tends to count every restriction proposed by capital as progressive. But whether a restriction is progressive is understood by looking at whom it loads the cost onto. A "slowdown" that makes open-weight models practically impossible, puts the small developer in an approval queue, and turns the large company's legal department into a competitive advantage accelerates monopolisation. It is not slowing itself that is classed; it is the architecture of the slowing.

The third: entering the alignment "West against China" or "China against the West." Both skip the same question — whose is the labour inside the model. To choose between two blocs is for the class being called to choose to abandon its own independent position.

Concrete suggestions

None of these will change a CEO's timetable. But they will strengthen your hand in the next "let's slow down" call, the next "safety" story, the next popular-science relay.

1. Ask every oversight proposal the three questions. Who chooses the auditor, who pays, where does the power of sanction come from? These three questions classify the brightest transparency pledge in a minute. You can ask the same three questions of OHS inspection in Turkey, of public tenders, of university ethics boards; the method is the same.

2. Turn the "pace" debate into a "right to stop" debate. Whoever is in front of you, let your question be this: who should hold the authority to stop this production? This question carries the debate from speed to authority, from authority to property. And the working class has a two-hundred-year answer to this question.

3. Address the computing worker. 1,386 signatures are the nucleus of a union. The tasks we described in The Computing Worker's Handbook are precisely for this stratum. International computing unions making contact with workers at the frontier laboratories, the computing union in Turkey taking up this agenda, is concrete and near work.

4. Write comments on popular-science relays — with a concept. Instead of writing "this is advertising" under a post, write "who pays the auditor." The first is a feeling, the second a method. And one in a hundred people who read that comment will, the next day, ask the same question themselves in another story.

5. Put Turkey's position into every debate. In this country AI is discussed not as a research programme but as an imported product. Into which branch, under which contract, under which supervision it enters — these are things to be spoken here, in the workplace, at the collective-agreement table and in the law. As with the KVKK debate, the headings of data and oversight are not distant from us; they are the nearest headings.

6. Keep demanding public and common infrastructure. Open model repositories, open data sets, public compute capacity, universities' independent evaluation capacity. The idea of an "independent evaluator" is a good idea; what is bad is that evaluator drawing a salary from the company they audit. The alternative is a public and autonomous scientific institution — and to demand this is not a utopia; it is a budget line.

7. Look at who joins a proposal and who does not. If a policy proposal is approved by rivals within hours, there is a common interest there; if there are those who do not join, the dividing line runs from there. Musk's three words, Altman's "we will do the same," and the silence of Meta and Nvidia are, on their own, a whole market map. This method of reading works in every sector: which employers' organisation says "yes" to a minimum-wage proposal, which stays silent?

8. Keep putting today's harm in front of tomorrow's apocalypse. Data-labelling labour, algorithmic management, the closing of junior posts, the production of the threat of unemployment, the spread of surveillance products. These are not probabilities; they are facts. As the debate slides toward apocalypse, bring the list back.

Dear Young Comrades,

On 12 September a CEO wrote that the sector must put its foot on the brake. An hour later a rival who had been in litigation with him for two years approved in three words; shortly afterwards the other rival said "we will do the same" and postponed his company's IPO. Where a two-year hostility is resolved in a day shows where that hostility in fact ended and where common interest began. This is the demand voiced three days ago in a resignation thread being translated, within three days, into capital's own language. The demand came from the worker; the solution was formulated at the owners' table; there was no worker at the table.

Let us say this plainly: there are true things inside this call. The risk is real, the need for coordination is real, the idea of independent evaluation is correct. But the correctness of an idea does not determine in whose hands, and for what work, that idea will serve. Independent oversight is oversight when the auditor is independent; a pacing agreement is an agreement when those who stand under the process that sets the pace are parties to it; safety is safety when those who demand safety are at the decision table. In all three the same thing is missing, and the name of the missing thing is plain.

And looking from Turkey we need to see a second thing as well: this debate comes to us together with the frame in which it is relayed. We cannot debate the content without stripping the frame. In Evrim Ağacı's post and in Bebar Bilim's videos there is the same gap: the subject is "we," "humanity," "artificial intelligence"; worker and owner, wage and profit, job and unemployment are not there. The job of popular-science channels is to defend science, and that job is valuable; but when relaying a capital group's policy text or product launch, the reflexes of the defence of science are not enough. What is needed there is a method that asks by whom, and inside which relation of interest, the data were produced. The name of that method is class analysis, and it is in no one's monopoly; it can be learned, it can be taught, it can be written as a comment under a post.

Young comrade, when you read the next "let's slow down" story the question you will ask is plain: who is putting their foot on the brake, whose foot is the brake under, and when it speeds up who pays the cost? No safety proposal that does not answer these three questions is our safety.

Comradely

Knowledge belongs to everyone.

Sources

Primary text

News and reactions

Camps and the open-weight debate

Axios, "AI's manifesto war", 2 August 2026 — Anthropic/Google DeepMind and Meta/Nvidia camps
The New Stack, "'Open weights are nowhere near a sufficient solution': Dario Amodei fires back on AI power" — open weights, the 10²⁶ FLOP threshold and SB 53
Fortune, "More than 1,200 AI workers are asking for Washington's help to build an AI slowdown plan", 29 July 2026

Criticisms

Paulo Carvão, "Should we 'Pace the Frontier?'", September 2026 — the lack of sanction in voluntary coordination
Curtis Pyke, "Dario Amodei's AI Slowdown: Safety or Regulatory Capture?", September 2026 — regulatory capture and open-weight models
StartupHub.ai, "Dario Amodei We Must Pace the Frontier Is Vague", 12 September 2026

The Hugging Face incident

Law and labour history

Sherman Antitrust Act (1890) and Loewe v. Lawlor ("Danbury Hatters", 1908) — the use of antitrust law against unions
Clayton Act (1914), section 6 — "The labor of a human being is not a commodity or article of commerce."
Occupational Health and Safety Law No. 6331, art. 13 — the right to refrain from work
ILO Occupational Safety and Health Convention No. 155 (1981)

Marxist frame

Karl Marx, Grundrisse, "The Fragment on Machines" — the general intellect
Karl Marx, Theses on Feuerbach (1845) — thought not independent of social relations; critique of idealism
Karl Marx, Capital vol. I, ch. 24 — accumulation imposed on the capitalist as an external coercive law of competition
V. I. Lenin, Imperialism, the Highest Stage of Capitalism (1916) — cartel, monopoly and intra-bloc coordination
Rudolf Hilferding, Finance Capital (1910) — cartelisation not abolishing competition but changing its form
Harry Braverman, Labor and Monopoly Capital (1974)
Albert O. Hirschman, Exit, Voice, and Loyalty (1970) — the distinction of exit and voice